When Annotation Patterns Are Learned: Skin-Tone Dependence in Dermatology Concepts
This paper audits SkinCon, a widely used dermatology concept dataset, for skin-tone-dependent annotation bias.
Within the same disease, erythema is labeled far less often on dark skin (odds ratio 0.17) while pigmentary concepts are labeled more often.
An evidential concept model trained on these annotations reproduces the bias, and three label-error detectors fall below a disease-only baseline at catching potential erythema omissions.
The work warns that concept bottleneck explanations for skin disease diagnosis cannot be trusted until annotations are audited within each subgroup.
The paper makes a timely and practically important argument that concept bottleneck models for dermatology risk inheriting tone-dependent annotation bias. The evidence for annotation asymmetry is statistically strong and consistent across deduplicated and full datasets, as well as under two sets of FST labels. However, the claim that detectors underperform is slightly understated for two of the three signals, which trail the baseline by modest margins while remaining above chance; only evidential uncertainty is near chance (AUC 0.48). The core limitation is that the paper never identifies individual mislabels with certainty, relying instead on aggregate odds ratios and underpowered caption checks.
The disease-adjusted odds ratio analysis is well-designed. By fitting $\operatorname{logit} P(C_{ik}=1) = \alpha_{k,d(i)} + \beta_k D_i$ within each disease, the authors convincingly rule out disease-mix confounding. The bootstrap percentile intervals are reported transparently, and the pattern replicates under the second FST label set and after deduplication. The demonstration that the concept model's predicted odds ratio ($e^{\beta'_k}$) tracks the annotation odds ratio is a crisp empirical result, and the clinical premise is grounded in established dermatology guidance.
A major concern is the granularity of evidence for mislabeling versus true biological differences. The authors acknowledge that hyperpigmentation is genuinely more common after inflammation in darker skin, but treat all deviation from parity as annotation dependence. For erythema, the external SkinCAP check is underpowered: the disease-adjusted SkinCAP odds ratio is 0.93 [0.16, 1.81] from only 109 images, so it neither confirms nor refutes true omissions. The paper's detector evaluation uses SkinCAP captions as ground truth, yet admits that dermatologists may have mentioned erythema because it is expected for the diagnosis rather than clearly visible. This circularity weakens the conclusion that the detectors are fundamentally broken rather than appropriately conservative.
Another concern is sample size and generalizability. After deduplication, only 242 dark-skin images remain. Subgroup AUCs in Table 1 have wide intervals that cross zero, and the model results come from a single architecture (evi-CEM). The claim about detectors is exploratory and would benefit from more than one concept model backbone.
The evidence strongly supports the existence of tone-dependent annotation patterns, but comparison to related work is thin. The paper cites confident learning, AUM, and evidential uncertainty, yet does not compare against recent instance-dependent or group-aware label-noise methods that might better handle image-dependent noise. The cited work by Wang et al. on fair classification with group-dependent label noise is mentioned only in passing; no empirical comparison is made. The single-model limitation also means we cannot tell whether the detector failure is specific to evidential uncertainty or general to concept models.
Reproducibility is mixed. The data sources are public (Fitzpatrick17k, SkinCon, SkinCAP), and the deduplication keep list from Abhishek et al. is referenced. However, no code or training configurations are provided. Hyperparameters for evi-CEM are only described as ResNet-34 backbone and five-fold cross-validation with three random seeds. The SkinCAP term search list for erythema is partially given, but the full caption processing pipeline (GPT-4 refinement, translation, dermatologist checking) lacks detail. Crucially, the keep list does not correct FST labels, and because the paper reports results on both deduplicated and full sets, reproduction requires careful tracking of which images were removed. Without released checkpoints or code, independent reproduction would be labor-intensive.
Concept bottleneck models (CBMs) let clinicians inspect and correct skin disease diagnoses through clinical concepts, but only if the concept annotations are reliable. We therefore audit the reliability of SkinCon, a widely used dermatology concept dataset, across skin tones. Within the same disease and with near duplicates removed, erythema is annotated far less often on dark than on light skin, with an odds ratio of 0.17. In contrast, hyperpigmentation and hypopigmentation are annotated more often, with odds ratios of 6.2 and 4.5, respectively. We refer to this pattern as tone-dependent concept annotation. An evidential concept model trained on these annotations learns the same pattern and reproduces it on unseen images. Such a model may agree with potential erythema omissions rather than contradict them. Consistent with this, three label error detectors underperform a disease-only baseline at finding erythema that dermatologist captions mention but SkinCon labels as absent. The three detectors are based on confident learning, training dynamics, and the model's own evidential uncertainty. Concept annotations should thus be audited within each subgroup before explanations built on them can be trusted.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.